Note: Descriptions are shown in the official language in which they were submitted.
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
SYSTEMS AND METHODS FOR AUTOMATIC MODEL GENERATION
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This application claims the benefit of U.S. Provisional Application
No.
62/940,113, filed 25-N0V-2019, which is incorporated herein in its entirety by
this
reference.
TECHNICAL FIELD
[0002] This invention relates to the data modeling field, and more
specifically to a
new and useful modelling system.
BACKGROUND
[0003] Data science tasks are typically performed by data scientists that
have
specialized knowledge related to generating, validating, and deploying machine
learning
models.
[0004] There is a need in the data modeling field to create new and useful
systems
and methods for data modeling. The embodiments of the present application
provide
such new and useful systems and methods.
BRIEF DESCRIPTION OF THE FIGURES
[0005] FIGURES 1A-B illustrate schematics of a system, in accordance with
embodiments.
[0006] FIGURES 2A-D illustrates a method, in accordance with embodiments.
[0007] FIGURE 3 illustrates schematics of a system, in accordance with
embodiments.
[0008] FIGURE 4 illustrates a method, in accordance with embodiments.
[0009] FIGURE 5 illustrates an exemplary user interface for receiving
selection of
a model purpose, in accordance with embodiments.
1
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
[0010] FIGURE 6 illustrates an exemplary user interface for selection of
a
generated model, in accordance with embodiments.
DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0011] The following description of preferred embodiments of the present
application are not intended to be limiting, but to enable any person skilled
in the art of
to make and use these embodiments described herein.
1. Overview
[0012] Data science tasks are typically performed by data scientists that
have
specialized knowledge related to data modeling. Such tasks often include
processing
raw data, feature selection, model generation, model validation, and model
execution.
[0013] Embodiments herein enable simplified data modeling by
automatically
generating a machine learning model based on supplied data.
[0014] In some variations, a model purpose for the model is identified,
and the
model is generated based on the identified purpose. In some variations, the
model
purpose is selected from a list of pre-determined model purposes by indication
of the
user interacting with a graphical user interface. For example, a user
interface can
display a list of selectable model purposes, and the system can receive user
selection of
one of the selectable model purposes via the user interface. In some
implementations,
the identified purpose is used to identify functional constraints of the model
that is to be
generated. For example, a "credit risk evaluation" purpose might identify a
first set of
constraints (e.g., features that are useful in predicting credit risk). In
some
implementations, the identified purpose identifies a specific domain (e.g,
"generic
lending product", "auto loan", "mortgage loan", "credit card", "installment
loan", etc.).
In some implementations, the system includes model purpose data that
identifies at
least one of the following for each model purpose supported by the system:
data sources,
data sets, features, canonical features, a prediction target, model type,
model
parameters, hyperparameters. In some implementations, the system includes
model
purpose data for an identified purpose, and the model purpose data is used to
generate
the model. For example, the model purpose data can be used to select features
or select
2
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
model parameters (type of model, target, hyperparameters, etc.). In some
implementations, model purpose data includes at least one model template.
[0015] In some variations, the model purpose data is generated by domain
experts (e.g., data scientists, business analysts, and the like) having
specific domain
knowledge related to the identified purpose. In some implementations, the
model
purpose data is received via a computing system (e.g., 131) (e.g., of a domain
expert).
For example, data scientists with experience with auto-loan originations can
generate
the model purpose data for an "auto loan origination" purpose, and this auto-
loan
origination model purpose data can be used to automatically generate models
for "auto
loan origination" purposes without further input from a data scientist.
[0016] In some variations, the model purpose relates to consumer loan
origination, and results of the model are used to determine whether to grant a
consumer
loan. In some variations, the model purpose relates to business loan
origination, and
results of the model are used to determine whether to grant a loan to a
business. In
other variations, the model purpose relates to loan repayment prediction, and
results of
the model are used to determine whether a loan already granted will be repaid.
In other
variations, the model purpose relates to identifying consumers to solicit for
a new loan,
and the results of the model are used to determine which consumers to solicit
to apply
for a loan. In other variations, the model purpose relates to identifying
curable loans,
and the results of the model are used to determine which consumers who are
delinquent on their loan payments are likely to cure if called. In some
variations the
model purpose relates to applicant identification, and results of the model
are used to
determine whether a consumer applying for a loan is a real person or a
synthetic
identity. In some variations, the model purpose relates to business loan
repayment, and
results of the model are used to determine whether a business applying for a
loan will
repay the loan. In some variations, the model purpose is further refined by
loan type,
including: retail loans such as mortgage loans, refis, home equity loans,
automotive
loans, RV loans, powersports loans, credit cards, personal loans, student
loans, and
commercial loans including equipment loans, revolving lines of credit,
accounts payable
financing, and other loan types, retail or commercial, without limitation.
3
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
[0017] Embodiments herein provide at least one of: automatic feature
selection,
automatic parameter selection, automatic model generation, automatic model
evaluation, automatic model documentation, automatic alternative model
selection,
automatic model comparison, automatic business analysis, automatic model
execution,
automatic model output explanation, and automatic model monitoring. In some
variations, a machine learning platform (e.g., a cloud-based Software as a
Service (SaaS)
platform) provides such features related to model generation, analysis and
validation
and deployment and monitoring. In some variations, an automatically generated
model
(e.g., generated by the machine learning platform) is compared with a pre-
existing
model (e.g., a model currently in use by a user of the platform, but not
generated by the
platform), and results of the comparison are provided to a user system. In
some
variations the comparison includes economic analysis describing the expected
business
outcomes likely to arise from deploying a new model.
[0018] In some implementations, loan data identifying loan attributes
(e.g., loan
amount, loan term, collateral value, collateral attributes), credit data used
to decide
whether to grant the loans (e.g., number of inquiries, number of
delinquencies, available
credit and utilization, credit bureau attributes, trended attributes, etc), a
credit policy,
and loan outcomes for the loans made previously (e.g., repaid successfully,
charged
off/unpaid, or delinquent for a given number of days), are used to project a
change in
business metrics (such as loan volume, new customers, revenue from interest,
loss rate,
loss amount, gross margin, and net profit) resulting from using a model
generated by
the system for a specific business application (e.g. granting auto loans,
credit line
increase, etc.). In some implementations, the system automatically generates
documentation, and the documentation identifies at least one of: selected
features,
reasons for choosing the selected features, how the model behaves in various
circumstances, business projections and the like.
[0019] In some variations, the system is a machine learning platform
(e.g., 110
shown in Figs. 1A-B). In some variations, the method includes at least one of:
accessing
data, detecting features, generating at least one model, evaluating at least
one model,
executing at least one model, generating explanation information for at least
one model,
4
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
generating business analysis for at least one model, generating monitors and
monitoring
outputs for at least one model, generating documentation information for at
least one
model, and providing documentation for at least one model.
2. System
[0020] In some variations, a system (e.g., 100 includes at least one of:
a feature
detection module (e.g., in), a feature selection module (e.g., 112), a model
generation
module (e.g., 113), a parameter selection module (e.g., 114), a model
evaluation module
(e.g., 115), a model selection module (e.g., 116), an output explanation
module (e.g., 117),
a model documentation module (e.g., 118), a user interface system (e.g., 119),
a model
execution module (e.g., 140, a model monitoring module (e.g., 141), and a data
store
(e.g., 150 that stores model purpose data.
[0021] In some variations, a system includes a machine learning platform
110. In
some variations, the machine learning platform is an on-premises system. In
some
variations, the machine learning platform is a cloud-system. In some
variations, the
machine learning platform functions to provide software as a service (SaaS).
In some
variations, the platform no is a multi-tenant platform. In some variations,
the platform
no is a single-tenant platform
[0022] In some implementations, the system no is a machine learning
platform
(e.g., no shown in Figs. 1A-B).
[0023] In some implementations, the system no includes at least one of:
the user
interface system 119 and the storage device 150. In some implementations, the
system
no includes at least one of the modules 111-118, 140 and 141 shown in Figs iA
and 1B.
[0024] In some implementations, at least one component (e.g., 111-119,
140, 141,
150 of the system no is implemented as program instructions that are stored by
the
system no (e.g., in storage medium 305, memory 322 shown in Fig. 3) and
executed by
a processor (e.g., 303A-N shown in Fig. 3) of the system no.
[0025] In some implementations, the system no is communicatively coupled
to at
least one data source (e.g., 121-123) via a network (e.g., a public network, a
private
network). In some implementations, the system no is communicatively coupled to
at
least one user system (e.g., 131) via a network (e.g., a public network, a
private network).
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
[0026] Fig. iB shows interactions of components of the system, in
accordance
with variations.
[0027] In some implementations, the storage device 150 stores model
purpose
data that identifies at least one of the following for each model purpose
supported by the
system: data sources, data sets, features, canonical features, a prediction
target, model
type, model parameters, hyperparameters. In some implementations, the storage
device
150 includes model purpose data for an identified purpose, and the model
purpose data
is used to generate a model. For example, the model purpose data can be used
to select
features or select model parameters (type of model, prediction target,
hyperparameters,
etc.). In some implementations, model purpose data includes at least one model
template. In some implementations, the template defines at least: canonical
features to
be used as model inputs; a model type; and a prediction target. In some
implementations, the template defines each model of an ensemble, and an
ensemble
function. In some implementations, the template defines, for at least one
model, input
sources. Input sources can be the feature detection module in, which provides
features
to the model. Input sources can also include an output of another model. For
example,
a first model can generate an output value that is used as an input of a
second model.
[0028] In some embodiments, the model purpose data is generated by domain
experts (e.g., data scientists) having specific domain knowledge related to
the identified
purpose. For example, data scientists with experience with auto-loans can
generate the
model purpose data for an "auto loan originations" purpose, and this auto-loan
model
purpose data can be used to automatically generate models for "auto loan
originations"
purposes without further input from a data scientist.
[0029] In some variations, the feature detection module in functions to
detect
features from accessed data (e.g., data provided by a user system, data
retrieved from a
data source, etc.). In some variations, the accessed data includes raw data.
In some
implementations, the feature detection module in receives the accessed data
via the
user interface system 119. In some implementations, the feature detection
module in
receives data from at least one of a loan management system (LMS) of the user
system
(e.g., 133), a loan origination system (LOS) of the user system (e.g., 132), a
data source
6
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
(e.g., 121-123) (e.g., TransUnion, Equifaxõ Schufa, LexisNexis, RiskView
credit bureau
data with full tradeline information, Experian, Clarity, a central bank,
Creditinfo,
Compuscan, etc.).
[0030] In some variations, at least one component of the system 110
generates
documentation information that documents processes performed by the component.
In
some variations, at least of the modules 111-118, 140 and 141 generates
documentation
information that describes processes performed by the module, and stores the
generated
documentation information in the model documentation module 118.
[0031] In some variations the documentation is based on analysis
performed on
the model (based on a model purpose, e.g., identified at S212 of the method
200 and
includes business analysis determined by the model purpose. For example in
auto
lending, the business reporting output includes business outcomes based on
switching
from an old model to a new model. In variants, business outcomes include the
projected
default rate for the new model (holding approval rate constant). In other
variants,
business outcomes include one or more of:the projected approval rate holding
risk
constant, a charge off amount projection; an interest income projection; and a
recovery
projection based on asset information and a depreciation formula. In some
variants, the
projected business outcomes from multiple model variations are compared and
summarized.
[0032] In some variations, the feature detection module in extracts
canonical
features from raw data accessed by the feature detection module 111. In some
implementations, each canonical feature is a semantically meaningful
representation of
information included in the accessed data. For example, the canonical feature
"Number
of Bankruptcies" can be extracted from raw data that includes features
"TransUnion
Count of Bankruptcies", "Experian Count of Bankruptcies", and "Equifax Count
of
Bankruptcies". In other words, rather than treating "TransUnion Count of
Bankruptcies", "Experian Count of Bankruptcies", and "Equifax Count of
Bankruptcies"
as individual features for purposes of model generation, the data from these
features are
used to determine a value for the canonical feature "Number of Bankruptcies".
7
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
[0033] In some implementations, the feature detection module 111 extracts
canonical features by applying predetermined transformation rules. In some
implementations the transformation rules are selected automatically based on
an
identified model purpose and properties of the model development data. In some
implementations, properties of the model development data are automatically
determined based on analysis methods and statistics such as: percent of
missing data,
min, max, median, mean mode, skew, variance, and other statistics without
limitation
overall and over time. In other implementations the transformation rules are
selected
based on metadata associated with each column in the training data. In some
implementations this metadata is computed based on predetermined rules. In
other
implementations the metadata is inferred based on statistics. For example if a
variable
with a low missing rate across 100,000 or more rows only takes on 5 distinct
numeric
values, the system (e.g, 100, 110) may infer that the variable is a
categorical and select a
transformation rule corresponding to "one hot" encoding, thereby generating a
series of
Boolean flags to replace the original low-cardinal values with it's numeric in
the
modeling data. In other implementations the transformation rules are selected
by
indication of the user within a graphical user interface (e.g., provided by
the user
interface system 119 shown in Fig. 1B).
[0034] In some implementations, the feature detection module in extracts
canonical features by performing any suitable machine learning process,
including one
or more of: supervised learning (e.g., using logistic regression, back
propagation neural
networks, random forests, decision trees, etc.), unsupervised learning (e.g.,
using an
Apriori algorithm, k-means clustering, etc.), semi-supervised learning,
reinforcement
learning (e.g., using a Q-learning algorithm, temporal difference learning,
etc.), and any
other suitable learning style. In some implementations, the feature detection
module in
implements any one or more of: a regression algorithm (e.g., ordinary least
squares,
logistic regression, stepwise regression, multivariate adaptive regression
splines, locally
estimated scatterplot smoothing, etc.), an instance-based method (e.g., k-
nearest
neighbor, learning vector quantization, self-organizing map, etc.), a
regularization
method (e.g., ridge regression, least absolute shrinkage and selection
operator, elastic
8
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
net, etc.), a decision tree learning method (e.g., classification and
regression tree,
iterative dichotomiser 3, C4.5, chi-squared automatic interaction detection,
decision
stump, random forest, multivariate adaptive regression splines, gradient
boosting
machines, etc.), a Bayesian method (e.g., naive Bayes, averaged one-dependence
estimators, Bayesian belief network, etc.), a kernel method (e.g., a support
vector
machine, a radial basis function, a linear discriminant analysis, etc.), a
clustering
method (e.g., k-means clustering, expectation maximization, etc.), an
associated rule
learning algorithm (e.g., an Apriori algorithm, an Eclat algorithm, etc.), an
artificial
neural network model (e.g., a Perceptron method, a back-propagation method, a
Hopfield network method, a self-organizing map method, a learning vector
quantization
method, etc.), a deep learning algorithm (e.g., a restricted Boltzmann
machine, a deep
belief network method, a convolutional network method, a stacked auto-encoder
method, etc.), a dimensionality reduction method (e.g., principal component
analysis,
partial least squares regression, Sammon mapping, multidimensional scaling,
projection
pursuit, etc.), an ensemble method (e.g., boosting, bootstrapped aggregation,
AdaBoost,
stacked generalization, gradient boosting machine method, random forest
method, etc.),
and any suitable form of machine learning algorithm. In some implementations,
the
feature detection module 111 can additionally or alternatively leverage: a
probabilistic
module, heuristic module, deterministic module, or any other suitable module
leveraging any other suitable computation method, machine learning method or
combination thereof. However, any suitable machine learning approach can
otherwise
be incorporated in the feature detection module in. Further, any suitable
model (e.g.,
machine learning, non-machine learning, etc.) can be used in detecting
canonical
features.
[0035] In some variations, the feature detection module in includes a
plurality of
feature detectors. In some variations, the feature detection module in
includes a
feature detector for each canonical feature.
[0036] In some variations, the feature detection module in detects all
canonical
features supported by the system 110. In some variations, the feature
detection module
in performs selective feature detection by detecting selected ones of the
canonical
9
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
features supported by the system 110. In some implementations, the feature
detection
module 111 selects canonical features for detection based on information
identifying a
model purpose. In some implementations, the feature detection module in
selects
canonical features for detection based on model purpose data associated with
an
identified model purpose. In some implementations, the feature detection
module in
selects canonical features for detection based on information received from a
feature
selection module (e.g., 112).
[0037] In some variations, the feature detection module in generates
training
data from data accessed by the feature detection module 111 (e.g., raw data,
data
provided by the user system, data retrieved from a data source, etc.). In some
variations
the feature detection module in automatically retrieves data from data sources
based
on information received from a user system (e.g., 131) via the user interface
system 119.
In some implementations, the information received by the feature detection
module in
from the user system via the user interface system 119 includes borrower
personal data
(name, address, government ID number), and information identifying selection
of a
model purpose. In some implementations, the feature detection module in
retrieves
training data records from various systems and data sources (e.g., 121-123)
automatically based on the data received from the user system . In some
implementations the data received from the user system includes borrower data
for a
sample of a population of user accounts identified by one or more of a
demographic
characteristic, an economic characteristic, and a credit characteristic.
[0038] In some implementations, the generated training data only includes
columns for canonical features detected by the feature detection module in,
and
respective values. In some implementations, the generated training data is
used by the
model generation module 113 to train a model (e.g., a model defined by model
purpose
data, e.g., 150, that corresponds to a model purpose identified by information
by a user
system, e.g., 131) during the model generation process. In some variations,
the feature
detection module in generates training data for a model template used by the
model
generation module 113, such that the training data includes at least one of:
data for
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
canonical features identified as inputs by the model template; and data for a
canonical
feature identified as a prediction target by the model template.
[0039] In some variations, the feature detection module 111 generates and
stores
documentation information that identifies at least one of: selected features,
data sources
accessed, time stamps for the accessed data, time stamps for detected
canonical
features, a description of the generated training data, data ranges,
statistical data related
to the detected features, name and description of of the transformation
applied to
generate the canonical feature, and the like.
[0040] In some variations, the user interface system 119 provides a
graphical user
interface (e.g., a web interface). In some variations, the graphical user
interface
includes a series of modules organized by business function, for example:
model
development, model adoption, and model operations. In some variations the
model
adoption module includes submodules including model risk, model compliance,
and
business impact. In some variations, the user interface system 119 provides a
programmatic interface (e.g., an application programming interface (API)) to
access
intermediate outputs, and final outputs from the system (e.g., no). In some
variations,
the user interface system 119 creates audit logs and reports that reflect
model variations
and detailed change logs. In some variations, the user interface system 119
provides
role-based access in which specific users only have access to certain modules.
In some
variations, the user interface system 119 is pre-integrated with other systems
such as
loan origination systems (LOS) (e.g., 132), data aggregators and credit
bureaus, so that
models can be developed, validated and published directly from the user
interface
system 119=. In this way, new model variations can be more easily tested and
deployed
where they can generate business impact. In some variations, the user
interface system
119 includes a monitoring dashboard that includes business impact monitoring,
model
monitoring, and system monitoring dashboards. In variants, the business impact
monitoring dashboard includes business metrics such as approval rate,
delinquency
rate, vintage loss curves, charge off value, interest income value, and
comparison to
prior models. In variants, the system 110 automatically gathers new data on
the
unfunded population in order to perform an automated ROT comparison between a
11
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
prior model and a new model based on the performance loans given by other
lenders to
the unfunded population.
[0041] In some variations, the feature selection module 112 functions to
select one
or more canonical features based on information identifying a model purpose.
In some
implementations, the feature selection module 112 receives the information
identifying a
model purpose from a user interface system (e.g., 119). In some variations,
the feature
selection module 112 selects one or more canonical features based on model
purpose
data associated with the identified model purpose.
[0042] In some variations, the feature selection model 112 incorporates
cost
information to select the set of data sources that deliver the maximum profit.
[0043] In some variations, the feature selection module 112 and the
parameter
selection module 114 are included in a selection module.
[0044] In some variations, the model generation module 113 generates at
least
one model based on information identifying a model purpose and the training
data (e.g.,
generated by the feature detection module, accessed from a data store,
accessed from a
data source, etc.). In some variations, the model generation module 113
generates at
least one model based on model purpose data (e.g., stored in 150) associated
with an
identified model purpose. In some variations, the model generation module 113
generates at least one model based on information (e.g., a model template)
received
from a parameter selection module (e.g., 114). In some variations, the model
generation
module 113 generates at least one model based on information received from a
feature
selection module (e.g., 112). In some implementations, the model purpose data
identifies a model template. In some implementations, each model template
defines a
model that uses canonical features detectable by the feature detection module
111. In
some implementations, the model generation model 113 generates a model that
uses
only canonical features detectable by the feature detection module 111. In
this manner,
generation of models can be constrained to models that use canonical features.
[0045] By virtue of using a feature detection module (e.g., in) that
processes raw
data to generate data in a canonical format, the format and identities of
canonical
features usable by the model generation module (e.g., 113) can be known in
advance,
12
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
thereby enabling the generation of model templates that can be used to
generate new
models.
[0046] In some variations, the model generation module 113 uses data
(training
data) output by the feature detection module 111 to train at least one model
generated by
the model generation module 113.
[0047] In some variations, the model generation module 113 functions to
generate
models using any suitable machine learning process, including one or more of:
supervised learning (e.g., using logistic regression, back propagation neural
networks,
random forests, decision trees, etc.), unsupervised learning (e.g., using an
Apriori
algorithm, k-means clustering, etc.), semi-supervised learning, reinforcement
learning
(e.g., using a Q-learning algorithm, temporal difference learning, etc.), and
any other
suitable learning style. In some implementations, generated models can
implement any
one or more of: a regression algorithm (e.g., ordinary least squares, logistic
regression,
stepwise regression, multivariate adaptive regression splines, locally
estimated
scatterplot smoothing, etc.), an instance-based method (e.g., k-nearest
neighbor,
learning vector quantization, self-organizing map, etc.), a regularization
method (e.g.,
ridge regression, least absolute shrinkage and selection operator, elastic
net, etc.), a
decision tree learning method (e.g., classification and regression tree,
iterative
dichotomiser 3, C4.5, chi-squared automatic interaction detection, decision
stump,
random forest, multivariate adaptive regression splines, gradient boosting
machines,
etc.), a Bayesian method (e.g., naive Bayes, averaged one-dependence
estimators,
Bayesian belief network, etc.), a kernel method (e.g., a support vector
machine, a radial
basis function, a linear discriminant analysis, etc.), a clustering method
(e.g., k-means
clustering, expectation maximization, etc.), an associated rule learning
algorithm (e.g.,
an Apriori algorithm, an Eclat algorithm, etc.), an artificial neural network
model (e.g., a
Perceptron method, a back-propagation method, a Hopfield network method, a
self-
organizing map method, a learning vector quantization method, etc.), a deep
learning
algorithm (e.g., a restricted Boltzmann machine, a deep belief network method,
a
convolutional network method, a stacked auto-encoder method, etc.), a
dimensionality
reduction method (e.g., principal component analysis, partial least squares
regression,
13
CA 03161968 2022-05-17
WO 2021/108553
PCT/US2020/062235
Sammon mapping, multidimensional scaling, projection pursuit, etc.), an
ensemble
method (e.g., boosting, bootstrapped aggregation, AdaBoost, stacked
generalization,
gradient boosting machine method, random forest method, etc.), and any
suitable form
of machine learning algorithm. In some implementations, a generated model can
additionally or alternatively leverage: a probabilistic module, heuristic
module,
deterministic module, or any other suitable module leveraging any other
suitable
computation method, machine learning method or combination thereof. However,
any
suitable machine learning approach can otherwise be incorporated in a
generated
model. Further, any suitable model (e.g., machine learning, non-machine
learning, etc.)
can be generated.
[0048] In
some variations, the feature selection module 112 functions to select
features to be detected by the feature detection module 111. In some
variations, the
feature selection module 112 functions to select features to be used by the
model
generation module 113. In some implementations, the feature selection module
112
selects features based on information identifying a model purpose (e.g.,
information
received via the user interface system 119). In some implementations, the
feature
selection module 112 selects features based on a model template that
identifies at least
one of input value features and prediction target features to be used during
model
generation.
[0049] In some variations, the parameter selection module 114
functions to
select parameters to be used during model generation (e.g., by the model
generation
module 113). In some implementations, the parameter selection module 114
selects
parameters based on information identifying a model purpose (e.g., information
received via the user interface system 119). In some implementations, the
parameter
selection module 114 selects parameters based on a model template that
identifies
parameters to be used during model generation. In some implementations, the
parameter selection module 114 selects at least one model template that
identifies
parameters to be used during model generation (e.g., by the model generation
module
113). In some implementations, parameters included at least one of: data
sources, data
14
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
sets, features, canonical features, a prediction target, model type, model
parameters,
and hyperparameters.
[0050] In variations, the parameter selection module 114 determines the
parameters used to train the model and the model generation module 113
produces a
model based on training data and the selected parameters. In some variations,
the
parameter selection module 114 enumerates various parameters and trains a
series of
models, then further selects the parameters that results in the maximum model
performance on a testing dataset. In variations, the model performance is
measured
based on AUC (area under the curve), max K-S and other statistics. In other
variations,
model performance is measured based on economic outcomes as determined by the
model purpose and an economic analysis method associated with the selected
purpose.
A search process for selecting model parameters can use any common search
method
such as grid search, bayesian search, and the like. The system (e.g., loo, no)
disclosed
herein conventional systems by making use of the model purpose to apply
economic
analysis to guide the feature selection process (performed by the feature
selection
module 112) and model parameter search process (performed by the parameter
selection module 114), which allows the system to produce and document models
that
yield higher economic performance (not just higher statistical performance).
In
lending, it is often the case that the economic consequence of a false
positive is different
than for false negatives. As such the disclosed system provides a new and
useful way of
incorporating this asymmetry into the model development process based on a
realistic
economic model corresponding to the specific model purpose (e.g., automotive
originations vs credit card originations). In an example, for the specific
purpose of auto
lending, a false negative could correspond to the case where the model
predicts the user
will repay when in fact they don't. In this case the cost to the lender is the
value of the
outstanding loan balance minus the value of the repossessed vehicle at auction
minus
costs. For bank cards (credit cards) there is no collateral (car to repossess)
and so in
embodiments, the economic consequences of a false negative are calculated
differently,
e.g., based on the outstanding balance, the cost of collections and the amount
collected.
Likewise in embodiments, for the model purpose of bankcard originations, the
value of a
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
true negative (repayment) might be based on the expected customer LTV
(interest
income over the average tenure and average bankcard balance for the proportion
of
customers that maintain balances in months). For auto loans the value of a
true
negative (repayment) might be based on the interest income for the one
specific loan.
These values can be used to produce weighted statistics such as a weighted F
measure
and a weighted AUC that incorporate the expected value of a true positive,
true negative,
false positive and false negative into the calculation vs assuming these are
valued
equally. Any suitable statistic may be used for this purpose. In this way,
during the
model development process, the parameter selection module 114 can incorporate
different expected values for true positives, true negatives, false positives
and false
negatives into the process of selecting the model parameters.
[0051] In some variations, the model documentation module 118 generates
model
documentation based on data stored by the model documentation module (and
optionally data received from other modules of the system 110 (e.g., 111-118,
140, 141).
In some implementations, the model documentation module 118 automatically
generates Model Risk Management (MRM) reports based on data received and/or
stored by the model documentation module 118.
[0052] In some variations, the model documentation module 118 stores
facts
about variables and features. In some variations, the model documentation
module 118
stores information that indicates the type of feature (numeric, categorical,
text, image),
where a variable came from (e.g., which database, which query, when
retrieved), which
variables contribute to a feature (e.g., average of which two variables,
maximum within
which column), how a feature was calculated (in human-readable language, e.g.,
English, and in computer-executable code), descriptive statistics,
visualizations, and
summaries, including: cardinality, histograms, distributions, analyses,
principal
components, anomalies, missing data, time-series, comparisons, a feature's
ideal value,
and protected class proxies (e.g., variables, features, or combinations of
variables and
features that can identify a protected class). In some variations, the model
documentation module 118 stores facts about the model development process,
including
who uploaded the data to develop the model, when it was uploaded, what changes
were
16
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
made to model inputs, parameters, and the like, by whom and when, comments
added
by model reviewers during the model review process, and other material
information
related to the model development process as orchestrated by a user interface.
[0053] In some variations, the model documentation module 118 stores
facts
about a model, including, without limitation: the training and validation data
sets, the
modeling method/machine learning algorithm used, the model tuning parameters,
model scores, model evaluation and analysis. In some variations, the model
documentation module 118 stores information that indicates lists of submodels
in an
ensembled model, model type, input feature list, and hyperparameters of a
model or
submodel, the parameter selection method and results, model performance
metrics,
feature contributions of a model or submodel. In some variations the feature
contributions are linked to the feature descriptions and descriptive
statistics and
metadata. In some variations, the model documentation module 118 stores
information
that indicates (for an ensemble model) an ensembling method, submodel, weights
of
submodels, and scoring functions for submodels and the scoring function for
the
ensemble. In some variations, the model documentation module 118 stores
information
related to the distribution of model scores and performance statistics overall
and by
segment. In other variations, the model documentation module 118 stores
information
about the feature contributions of the ensemble. In some variations, the model
documentation module 118 includes a knowledge repository, as described in U.S.
Patent
Application No. 16/394,651 ("SYSTEMS AND METHODS FOR ENRICHING
MODELING TOOLS AND INFRASTRUCTURE WITH SEMANTICS"), filed 25-APR-
2019, the contents of which is incorporated herein.
[0054] In some variations, the model evaluation module 115 functions to
evaluate
at least one model generated by the model generation module 113. In some
variations,
the model evaluation module 115 performs accuracy analysis for at least one
model
generated by the model generation module 113. In some variations, the accuracy
analysis includes computing a max K-S, Gini coefficient, or AUC statistic on a
test data
set. In some variations, the test data set is an out-of-time hold-out data set
( a data set
from a period after the model development data in time). In some variations
the model
17
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
evaluation module 115 calculates statistics on subsets of the test data, for
example, K-S
and AUC by day, week, month. In some variations, dispersion metrics are
calculated
for these accuracy metrics over time, such as the variance in AUC week over
week. In
some variations, the model evaluation module 115 performs economic analysis
comparing a model with another model or method and estimating the economic
impact
of adopting a new model based on the model purpose (as described herein with
respect
to the parameter selection module 114). In some variations, the model
evaluation
module 115 performs fair lending disparate impact analysis for at least one
model
generated by the model generation module 113. In some variations, the model
evaluation module 115 performs fair lending disparate impact analysis using a
method
described in U.S. Patent Application No. U.S. Application No. 16/822,908
("SYSTEMS
AND METHODS FOR MODEL FAIRNESS"), filed 18-MAR-2020, the contents of which
is incorporated herein. In some variations, the evaluation module 115 stores
evaluation
results in the model documentation module 118.
[0055] In some variations, the model selection module 116 selects at
least one
model generated by the generation module 113, based on results of the model
evaluation
module 115. For example, the generation module 113 can generate several
models, the
evaluation module can evaluate each model based on fair lending disparate
impact
analysis,accuracy analysis, and economic impact analysis, and the selection
module 116
can select a model that satisfies constraints for economics, accuracy and
fairness (e.g.,
constraints provided via the user interface system 119). In some variations,
the model
selection module 116 stores selection results (and optionally a rationale for
a selection,
e.g., economics, accuracy and fairness analysis results used in the selection)
in the
model documentation module 118.
[0056] In some variations, the model execution module 140 functions to
execute
at least one model generated by the model generation module 113. In some
variations,
the model execution module 140 executes at least one model generated by the
model
generation module 113 by using data output by the feature detection module 111
as input
data. In some implementations, each model executed by the model execution
module
140 receives input data from the feature detection module 111. In this manner,
the
18
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
feature detection module 111 performs pre-processing of raw data used during
model
execution. In some variations, during model execution, raw input data is
received by
the feature detection module in, the feature detection module in processes the
raw
data, and this processed data is provided as input to the model (or models)
being
executed by the model execution module 140.
[0057] In some variations, the output explanation module 117 functions to
generate explanation information for output generated by a model being
executed by the
model execution module 140. In some variations, the output explanation module
117
functions to generate explanation information by performing a method described
in U.S.
Patent Application No. 16/297,099, filed 8-MAR-2019, entitled "SYSTEMS AND
METHODS FOR PROVIDING MACHINE LEARNING MODEL EXPLAINABILITY
INFORMATION BY USING DECOMPOSITION", by Douglas C. Merrill et al, the
contents of which is incorporated herein. In some variations, the output
explanation
module 117 functions to generate explanation information by performing a
method
described in U.S. Patent Application No. 16/688,789 ("SYSTEMS AND METHODS FOR
DECOMPOSITION OF DIFFERENTIABLE AND NON-DIFFERENTIABLE MODELS"),
filed 19-NOV-2019, the contents of which is incorporated by reference. In some
variations, the output explanation module 117 functions to generate
explanation
information by performing a method described in U.S. Patent Application No.
U.S.
Application No. 16/822,908 ("SYSTEMS AND METHODS FOR MODEL FAIRNESS"),
filed 18-MAR-2020, the contents of which is incorporated herein.
[0058] In some variations, the explanation module 117 generates FCRA
Adverse
Action Reason Codes for output generated by a model being executed by the
model
execution module 140.
[0059] In some variations, the monitoring module 141 functions to monitor
performance of at least one model in production. In some variations, the
monitoring
module 141 monitors by performing a method described in U.S. Patent
Application No.
16/394,651 ("SYSTEMS AND METHODS FOR ENRICHING MODELING TOOLS AND
INFRASTRUCTURE WITH SEMANTICS"), filed 25-APR-2019, the contents of which is
incorporated herein. In some variations, the monitoring module 141 performs
19
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
monitoring based on at least one of: data stored by the documentation module
118, data
provided by the execution module 140, and data provided by the explanation
module
117.
[0060] In some variations, the monitoring module 141 functions to monitor
the
economic performance of at least one model in production. In variations,
economic
performance is computed based on the model purpose and performance data
gathered
from the customer's systems and includes approval rate, projected default
rate,
projected losses, projected profits, actual default rate, actual losses, and
actual profits.
In other variations, economic performance monitoring includes calculating
counterfactual scenarios considering what would have happened to if the
customer had
left their original model in production. In variants, the method of
calculating
counterfactual economic scenarios for models with loan origination purposes
includes
retrieving data from credit bureaus and other data sources about applications
for loans
that were rejected by the new model but that would have been accepted by an
old model.
Other counterfactual economic analysis methods are employed for models with
different
purposes. In this way the monitoring method disclosed herein improves upon the
state
of the art by incorporating knowledge of the model purpose and data collected
during
the model development and evaluation process to produce meaningful business
results
monitoring outputs for the plurality of model purposes the system supports.
3. Method
[0061] As shown in Fig 2A, a method 200 includes at least one of:
accessing data
(S21o); detecting features (S22o); generating at least one model (S23o);
evaluating at
least one model (S24o); executing at least one model (S25o); generating
business
analysis information (S26o); generating explanation information for at least
one model
(S27o); monitoring at least one model (S28o); and generating documentation
information for at least one model (S29o). Fig. 4 shows a schematic
representation of
an implementation of the method 200.
[0062] In some variants, at least one component of the system loo
performs at
least a portion of the method 200.
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
[0063] In some variations, the machine learning platform 110 performs at
least a
portion of the method 200. In some variations, at least one component of the
system
no performs at least a portion of the method 200.
[0064] In some implementations, a cloud-based system performs at least a
portion of the method 200. In some implementations, a local device performs at
least a
portion of the method 200.
[0065] In some variations, accessing data S210 functions to access data
from at
least one of a user system (e.g., 131-133) and a data source (e.g., 121-123)
that is external
to a user system (e.g., a credit bureau system, etc.). In some variations, the
feature
detection module in performs at least a portion of S210. In some variations,
the user
interface system 119 performs at least a portion of S210.
[0066] Accessing data S210 can include at least one of: accessing user
dataS2n,
identifying a purpose S212 and generating documentation information S213 shown
in
Fig. 2B.
[0067] Accessing user data S211 can include accessing user data from a
user
system (e.g., 131-133 shown in Fig. 1B), or a data source identified by a user
system.
[0068] Identifying a purpose S212 functions to identify a purpose for a
model to
be generated by the system (e.g., 110). In some variations, the system no
(e.g., user
interface system 119) identifies the purpose from information provided by a
user system
(e.g., 131). In variants, the system 110 receives information identifying user
selection of
a model purpose via a user interface system (e.g., 119). Fig. 5 shows an
exemplary user
interface that receives user input for a model purpose ("Model Type", "Product
Line").
In some variations, the system no identifies the purpose by processing data
used to
generate a model (e.g., training data). For example, a system no can receive
data from
a Loan Origination System (e.g., 132), and process the received data to
identify a model
purpose. The Loan Origination Data can identify the data as being data for an
auto
loan, and the system no can automatically identify the model purpose as "auto
loan".
For example, the data can include data that identifies a car that is subject
to the loan,
and this information can be used to infer that the data relates to an "auto
loan".
21
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
However, any suitable process for identifying a model purpose can be performed
by the
system 110.
[0069] In variants, identifying a purpose at S212 includes accessing
model
purpose data that is stored in association with the identified model purpose.
In some
implementations, the model purpose data is accessed (directly or indirectly)
from the
model purpose data store (e.g., 150.
[0070] Generating documentation S213 functions to generate documentation
information related to processes performed during S210. In variants, the
documentation information is managed by the model documentation module 118.
[0071] In some variations, detecting features S220 includes generating
training
data from the data accessed at S210. In some variations, detecting features
S220
includes detecting features, and generating training data that includes the
detected
features. In some variations, the feature detection module in performs at
least a
portion of S220.
[0072] Detecting features S220 can include at least one of: selecting
features
S221, detecting canonical features from accessed data S222, and generating
documentation information S223, as shown in Fig. 2C.
[0073] Selecting features S221 functions to select features to be
detected by the
system 110 (e.g., by using the feature detection module in). In some
implementations,
the feature selection module 112 performs feature selection, as described
herein with
respect to the feature selection module 112. In some implementations,
canonical
features are selected at S221. In some implementations, the features are
selected (e.g.,
by the feature selection module 112) based on model purpose data (e.g., stored
in 150
associated with the purpose identified at S212. In some implementations, the
model
purpose data includes a model template, as described herein.
[0074] Detecting canonical features S222 functions to detect at least one
canonical feature from data accessed at S210. In some variations, the feature
detection
module in performs S222 (as described herein with respect to the feature
detection
module in). In some variations, S222 includes detecting canonical features
selected at
S221. In some variations, S222 includes detecting only canonical features
selected at
22
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
S221. In some implementations, a plurality of feature detectors are used to
perform
S222. In some variations, S222 includes generating training data from the
detected
canonical features.
[0075] Generating documentation information at S223 functions to generate
documentation information related to processes performed during S220. In some
implementations, the documentation information is managed by the model
documentation module 118.
[0076] Generating a model S23o can include at least one of: selecting a
model
type S231, generating a model based on detected features S232, selecting
parameters
S233, and generating documentation information related to model generation
S234, as
shown in Fig. 2D. In some variations, the model generation module 113 performs
at
least a portion of S23o.
[0077] Generating a model S23o can include: generating a model based on
parameters identified by model purpose data (e.g., a model template) (e.g.,
stored in
150) associated with the purpose identified at S212; and training the model by
using
training data generated at S220.
[0078] In some variations, selecting a model type at S231 includes
selecting the
model type based on model purpose data (e.g., a model template) (e.g., stored
in 150).
[0079] In some variations, generating the model based on detected
features S232
includes defining the model to include as input features, only features
detectable by the
feature detection module 111. In some variations, S232 includes defining the
model to
include as a prediction target, only features detectable by the feature
detection module
111.
[0080] In some variations, selecting model parameters S233 includes
selecting at
least one of hyperparameters, feature weights, and the like. In some
variations, the
model parameters are selected based on model purpose data (e.g., a model
template)
(e.g., stored in i5o). In some variations, the model parameters are selected
based on
model economic analysis methods associated with the model purpose data (e.g.,
stored
in i5o). In an example, model purpose data identifies, for at least one model
purpose,
model parameters associated with economic analysis methods that will be
performed for
23
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
the model generated for the model purpose. For example, for an auto loan
origination
purpose, the model purpose data identifies model parameters that enable
business
analysis related to auto loan origination.
[0081] Generating documentation information at S234 functions to generate
documentation information related to processes performed during S23o. In some
implementations, the generated documentation information is managed by the
model
documentation module n8.
[0082] In variants, the model(s) generated at S23o can be any suitable
type of
model. Models generated at S23o can include differentiable models, non-
differentiable
models, and ensembles (which can include any combination of differentiable and
non-
differentiable models, ensembled using any suitable ensembling function).
[0083] In a first example, a model generated at S23o includes a gradient
boosted
tree forest model (GBM) that outputs base scores by processing base input
signals.
[0084] In a second example, a model generated at S23o includes a gradient
boosted tree forest model that generates output by processing base input
signals. The
output of the GMB is processed by a smoothed Empirical Cumulative Distribution
Function (ECDF), and the output of the smoothed ECDF is provided as the model
output (percentile score).
[0085] In a third example, a model generated at S23o includes sub-models
(e.g., a
gradient boosted tree forest model, a neural network, and an extremely random
forest
model) that each generate outputs from base input signals. The outputs of each
sub-
model are ensembled by using a linear stacking function to produce a model
output
(percentile score).
[0086] In a fourth example, a model generated at S23o includes sub-models
(e.g.,
a gradient boosted tree forest model, a neural network, and an extremely
random forest
model) that each generate outputs from base input signals. The outputs of each
sub-
model are ensembled by using a linear stacking function. The output of the
linear
stacking function is processed by a smoothed ECDF, and the output of the
smoothed
ECDF is provided as the model output (percentile score).
24
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
[0087] In a fifth example, a model generated at S23o includes sub-models
(e.g., a
gradient boosted tree forest model, and a neural network) that each generate
outputs
from base input signals. The outputs of each sub-model (and the base signals
themselves) are ensembled by using a deep stacking neural network. The output
of the
deep stacking neural network is processed by a smoothed ECDF, and the output
of the
smoothed ECDF is provided as the model output (percentile score).
[0088] However, the model can be any suitable type of model, and can
include
any suitable sub-models arranged in any suitable configuration, with any
suitable
ensembling and other processing functions.
[0089] Evaluating the model S24o functions to evaluate a model generated
at
S23o, generate evaluation information for the model. In some variations, the
model
evaluation module 115 performs at least a portion of S24o.
[0090] In some variations, evaluating the model at S24o includes
performing
accuracy analysis for at least one model generated at S23o, as described
herein. In
variants, the evaluation information includes results of the accuracy
analysis.
[0091] In some variations, evaluating a model (S24o) includes generating
economic analysis information for at least one model generated at S23o. In
some
variations, the economic analysis information is generated based on the model
purpose
and a comparison of models or methods. In some variations, generating the
economic
analysis information includes computing a value for at least one business
metric for the
model generated at S23o. In some implementations, the model purpose data
(accessed
at S212) defines the each business metric associated with the model purpose,
and values
for these business metrics are computed (at S24o) for the model generated at
S23o. In
some implementations, a value for at least one business metric is also
computed for an
original model used for the purpose identified at S212. In some
implementations,
business metric values for the original model are compared with corresponding
business
metric values for a model generated at S23o. In some implementations, results
of the
comparison between the business metric values for the original model and the
business
metric values for the model generated at S23o are included in the generated
economic
analysis information.
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
[0092] In an example, performing economic analysis at S24o includes
generating
economic analysis information identifying projected values for business
metrics for a
deployed instance of a model generated at S23o. Example business metrics
projected at
S24o include one or more of: loan volume, new customers, customer acquisition
cost,
revenue from interest, loss rate, loss amount, gross margin, and net profit.
[0093] For example in auto lending, the business reporting output
includes
business outcomes based on switching from an old model to a new model. In
variants,
business outcomes include the projected default rate for the new model
(holding
approval rate constant). In other variants, business outcomes include one or
more of:
the projected approval rate holding risk constant; a charge off amount
projection; an
interest income projection; and a recovery projection based on asset
information and a
depreciation formula. In some variants, the projected business outcomes from
multiple
model variations are compared and documented.
[0094] In some variations, evaluating a model at S24o includes performing
fair
lending disparate impact analysis for at least one model generated at S23o, as
described
herein. In variants, the evaluation information includes results of the fair
lending
disparate impact analysis and includes fairness metrics and business outcomes
under
various scenarios. The scenarios help the user choose which model to select
and
document the reasons for their selection via a user interface (e.g., 119).
[0095] In some variations, evaluating a model S24o includes selecting
(e.g., by
using the model selection module 116) at least one model generated at S23o,
based on
model evaluation results generated at S24o. Fig. 6 shows an exemplary user
interface
for selecting a model ("Auto 2020 Version 2"), based on model evaluation
results
("Accuracy", "Fairness", "Savings (Loss Reduction)") generated at S24o.
[0096] In some variations, evaluating a model at S24o includes generating
documentation information related to processes performed during S24o. In
variants,
the documentation includes the generated evaluation information. In some
implementations, the documentation information is managed by the model
documentation module 118.
26
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
[0097] Executing a model at S25o functions to execute a model generated
at
S23o. In some variations, the model execution module 140 performs at least a
portion
of S25o. In some variations, S25o includes executing at least one model
generated at
S23o. In some variations, S25o includes executing at least one model generated
by the
model generation module 113 by using data output by the feature detection
module in
as input data. In some implementations, each model executed at S25o receives
input
data from the feature detection module 111. In this manner, the feature
detection
module in performs pre-processing of raw data used during model execution (at
S25o).
In some variations, during model execution, raw input data is received by the
feature
detection module in, the feature detection module in processes the raw data,
and this
processed data is provided as input to the model (or models) being executed at
S25o.
[0098] In some variations, S25o includes generating at least one model
output by
using at least one model generated at S23o. In variations, S25o includes
generating
model outputs for the purposes of validating the model outcomes for a user-
specified
scenario, such as a change in applicant specified by the user via a user
interface.
[0099] In some variations, S25o includes generating documentation
information
related to processes performed during S25o. In some implementations, the
documentation information is managed by the model documentation module 118.
[00100] Generating business analysis information at S26o functions to
generate
business analysis information by using model output generated by the deployed
model
(e.g., at S25o). In variants, generating business analysis information
includes
identifying one or more of: approval rate, delinquency rate, vintage loss
curves, charge
off value, and interest income value, related to loans originated by using
output
generated by the deployed model (or models) at S25o. In variants, model
purpose
information (accessed at S212) defines at least one business analysis process,
and the
system (e.g., no) generates the business analysis information system (at S26o)
by
performing at least one business analysis process defined by the accessed
model
purpose information. In this manner, business analysis is performed in
accordance with
the identified model purpose (identified at S212), and business analysis can
be tailored
to a specific model purpose. In variants, the user provides business analysis
inputs via a
27
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
user interface. The system provides good default values for business inputs
based on the
business purpose and the model development data, based on a set of
predetermined
rules or a model. The user can modify the default values for business inputs
based on
their specific business circumstances, for example, by providing an average
total cost of
a loan default, an average interest income, a customer lifetime value, and
other values
and costs that enter into the calculation of various business metrics such as
profitability.
The documentation model reflects the method and assumptions selected by the
user in
the documentation.
[00101] S27o functions to generate explanation information for model
output
generated at S25o. In some variations, the output explanation module 117
performs at
least a portion of S27o. In some variations, S27o includes generating
explanation
information as described herein with respect to the output explanation module
117. In
some variations, S27o includes generating FCRA Adverse Action Reason Codes for
output generated at S25o. In some variations S26o includes generating FCRA
Adverse
Action Reason Codes for output generated at S25o based on a mapping from
individual
input features to more general reason codes and aggregating contribution of
individual
input features belonging to the same reason code.
[00102] In some variations, S27o includes generating documentation
information
related to processes performed during S27o. In some implementations, the
documentation information is managed by the model documentation module 118.
[00103] S28o functions to monitor at least one model being executed at
S25o. In
some variations, the model monitoring module 141 performs at least a portion
of S280.
In some variations, S28o includes monitoring performance of at least one model
in
production, as described herein with respect to the monitoring module 141. In
some
variations, S28o functions to detect at least one of feature drift, unexpected
inputs,
unexpected outputs, population stability, unexpected economic performance, and
the
like. In some variations, S28o functions provide an alert to at least one
system (e.g.,
131-133 shown in Fig. 1B) in response to detecting at least one of feature
drift,
unexpected inputs, unexpected outputs, population stability, economic
performance,
and the like. In variations, S28o assesses the importance of monitoring
outputs based
28
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
on properties of the model development data and a model purpose. In some
variations
the criteria for assessing the importance of monitoring outputs is based on a
model. In
variations, the importance assessment is used to determine whether to send an
alert to a
user indicating an important monitoring output was generated that warrants
further
attention. In this way, the user may take corrective action when a high
incidence of
feature drift, or unexpected economic performance occurs, for example by
rebuilding
the model based on new data or observations. In variations, an alert leads to
a user
interface that guides the user through a process to remediate the conditions
causing the
alert. In variations, this process is configured based on a model purpose,
properties of
the model development data and business analysis inputs.
[00104] In some variations, generating documentation at S290 includes
providing
at least a portion of the document information generated during performance of
the
method 200 (e.g., at S210, S220, S230, S240, S250, S260 and S270). In some
implementations, the documentation includes evaluation information generated
at
S240. In some implementations, the documentation includes business analysis
information generated at S250. In some implementations, the documentation
includes
explanation information generated at S270. In some implementations, the
documentation includes monitoring information generated at S280. In some
variations, the model documentation module 118 performs at least a portion of
S290. In
some variations, the user interface system 119 performs at least a portion of
S290. In
some variations, S290 functions to provide a Model Risk Management (MRM)
report to
a user system (e.g., 131).
[00105] In some variations, the user interface system 119 provides the
user system
131 with information identifying loan origination costs and profits resulting
from loan
generation and management by using an existing system or process of the user
system
131, and information identifying loan origination costs and profits predicted
by using a
model generated by the system 110. For example, the system 110 can access loan
origination data (and related data) from the user system (e.g., from the LOS
132),
identify actual losses from loan defaults, determine whether the model
generated by the
system 110 would have approved the loans resulting in actual losses, and
determine a
29
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
predicted loan default loss that would have been realized had the model
(generated by
the system 110) been used to approve the loans processed by the user system.
In this
manner, an entity managing the user system can learn whether use of the model
generated by the platform 110 would have reduced loan default losses. As
another
example, the system 110 can identify loan applications that were denied by the
entity but
would have been approved by using the model, and predict profits and defaults
associated with approving these loans. In this manner, the entity can learn
whether the
model can be used to approve more loans (resulting in increased profit), while
at the
same time managing default risk, thereby resulting in increased profit.
[00106] In variations, the user interface system 119 provides functions
that enable
model risk and compliance teams to comment on the Model Risk Management Report
and provide written feedback which is recorded and categorized by severity,
and
automatically routed to the user that is preparing the model for review. In
variations,
this feedback is further captured and managed in the model documentation
module 118.
In variations, a model review process is facilitated, in which multiple
stakeholders
review the model and provide feedback for the user preparing the model for
review. In
other variations, this feedback is used to modify the model. In some
variations, the user
interface system 119 facilitates model modifications including dropping an
input feature,
adding a monotonicity constraint, selecting different training data, modifying
an
adverse action reason code mapping, and the like. Such model modifications are
again
reflected in the model documentation module 118 and in the model
documentation.
[00107] In some variations, the system no is implemented by one or more
hardware devices. Fig. 3 shows a schematic representation of architecture of
an
exemplary hardware device 300.
[00108] In some variations, a hardware device (e.g., 300 shown in Fig. 3)
implementing the system 110 includes a bus 301 that interfaces with the
processors
303A-N, the main memory 322 (e.g., a random access memory (RAM)), a read only
memory (ROM) 304, a processor-readable storage medium 305, and a network
device
311. In some variations, the bus 301 interfaces with at least one of a display
device 391
and a user input device 381.
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
[00109] In some variations, the processors 303A-303N include one or more
of an
ARM processor, an X86 processor, a GPU (Graphics Processing Unit), a tensor
processing unit (TPU), and the like. In some variations, at least one of the
processors
includes at least one arithmetic logic unit (ALU) that supports a SIMD (Single
Instruction Multiple Data) system that provides native support for multiply
and
accumulate operations.
[00110] In some variations, at least one of a central processing unit
(processor), a
GPU, and a multi-processor unit (MPU) is included.
[00111] In some variations, the processors and the main memory form a
processing unit 399. In some variations, the processing unit includes one or
more
processors communicatively coupled to one or more of a RAM, ROM, and machine-
readable storage medium; the one or more processors of the processing unit
receive
instructions stored by the one or more of a RAM, ROM, and machine-readable
storage
medium via a bus; and the one or more processors execute the received
instructions. In
some embodiments, the processing unit is an ASIC (Application-Specific
Integrated
Circuit). In some embodiments, the processing unit is a SoC (System-on-Chip).
[00112] In some variations, the processing unit includes at least one
arithmetic
logic unit (ALU) that supports a SIMD (Single Instruction Multiple Data)
system that
provides native support for multiply and accumulate operations. In some
variations the
processing unit is a Central Processing Unit such as an Intel processor.
[00113] In some variations, the network adapter device 311 provides one or
more
wired or wireless interfaces for exchanging data and commands. Such wired and
wireless interfaces include, for example, a universal serial bus (USB)
interface,
Bluetooth interface, Wi-Fi interface, Ethernet interface, near field
communication
(NFC) interface, and the like.
[00114] Machine-executable instructions in software programs (such as an
operating system, application programs, and device drivers) are loaded into
the memory
(of the processing unit) from the processor-readable storage medium, the ROM
or any
other storage location. During execution of these software programs, the
respective
machine-executable instructions are accessed by at least one of processors (of
the
31
CA 03161968 2022-05-17
WO 2021/108553 PCT/US2020/062235
processing unit) via the bus, and then executed by at least one of processors.
Data used
by the software programs are also stored in the memory, and such data is
accessed by at
least one of processors during execution of the machine-executable
instructions of the
software programs. The processor-readable storage medium is one of (or a
combination of two or more of) a hard drive, a flash drive, a DVD, a CD, an
optical disk,
a floppy disk, a flash storage, a solid state drive, a ROM, an EEPROM, an
electronic
circuit, a semiconductor memory device, and the like.
[00115] The system and methods of the preferred embodiments and variations
thereof can be embodied and/or implemented at least in part as a machine
configured to
receive a computer-readable medium storing computer-readable instructions. In
some
variations, the instructions are executed by computer-executable components
integrated
with the system and one or more portions of the processor and/or the
controller. The
computer-readable medium can be stored on any suitable computer-readable media
such as RAMs, ROMs, flash memory, EEPROMs, optical devices (CD or DVD), hard
drives, floppy drives, or any suitable device. In some variations, the
computer-
executable component is a general or application specific processor, but any
suitable
dedicated hardware or hardware/firmware combination device can alternatively
or
additionally execute the instructions.
[00116] Although omitted for conciseness, the preferred embodiments
include
every combination and permutation of the various system components and the
various
method processes.
[00117] As a person skilled in the art will recognize from the previous
detailed
description and from the figures and claims, modifications and changes can be
made to
the preferred embodiments of the invention without departing from the scope of
this
invention defined in the following claims.
32
